When the Claim About AI Is the Failure

What the FTC has acted on since 2025

An “AI” service that did not listen. According to the FTC’s May 2026 complaints, Cox Media Group, MindSift and 1010 Digital Works told customers that their “Active Listening” service used conversations overheard by smart devices to target local ads, and that consumers had opted in. The FTC said the service did not listen to conversations or use voice data; it resold, at a significant markup, email lists bought from data brokers. The final orders of August 27, 2026 require $930,000 in total for customer redress and bar the firms from misrepresenting their services. The FTC’s consumer protection director said the product did not do what the companies claimed it did.

An accuracy figure that did not hold. In its April 2025 complaint, the FTC said Workado claimed 98.3 percent accuracy for its AI content detector, which it marketed mainly for marketing and other plain-language text, relying only on its model developers’ tests on academic writing. On non-academic text, the developers’ best overall accuracy was 74.5 percent, and the model correctly flagged AI-written text only 53.2 percent of the time. The final order, approved in August 2025, bars any accuracy claim unless the company has competent and reliable evidence each time it makes one, and requires it to keep that evidence.

An “AI lawyer” never tested against a lawyer. DoNotPay sold its service as “the world’s first robot lawyer”; according to the FTC’s complaint, it never tested whether its “AI lawyer” performed at the level of a human lawyer and did not hire attorneys to check its legal features. The February 2025 final order requires $193,000 in monetary relief and bars it from claiming its service performs like a lawyer without sufficient evidence.

Accuracy and bias claims without testing. In January 2025 the FTC finalized an order against IntelliVision over allegations that it had no evidence for claims that its AI facial recognition had one of the highest accuracy rates on the market and zero gender or racial bias, and that it had not trained the software on millions of faces as claimed. The order bars accuracy and bias claims without competent and reliable testing. In April 2025 accessiBe agreed to pay $1 million over claims that its plug-in could make any website meet accessibility guidelines.

Earnings promised from AI. Air AI sold what it called conversational AI that could replace customer-service staff, with promises of earnings and refunds, according to the FTC’s complaint. In March 2026 a federal court in Arizona entered an order with an $18,000,000 judgment, with the remainder suspended once $50,000 is paid, and a permanent ban on selling business opportunities; the defendants neither admitted nor denied the allegations. The order allows earnings claims only with written substantiation in hand when the claim is made. The same pattern ran through other 2025 cases over AI-powered “passive income” schemes.

What the FTC is not doing. In December 2025 the FTC set aside its 2024 order against Rytr, an AI writing tool, finding the complaint failed the legal requirements of the FTC Act and that the order unduly burdened AI innovation. It said it “will continue to hold accountable actors that use AI to violate the law or deceive consumers about the capabilities of their generative AI”. Its April 2026 testimony to the Senate describes the approach as targeting deception rather than regulating the technology. In July 2026 it proposed a policy statement on AI companies that distort their systems’ outputs for undisclosed ideological objectives; comments closed on July 31, 2026, and the FTC still lists the statement as proposed. The line is the claim, not the tool.

Securities regulators and states

Public companies. In January 2025 the SEC found that Presto Automation, a restaurant-technology company listed on Nasdaq until September 2024, described third-party speech recognition in its filings as its own technology, and claimed its voice AI eliminated the need for human order taking when people abroad handled most orders. The SEC also found that, for more than two years, no one at Presto was formally responsible for the accuracy of its filings. Without admitting or denying the findings, Presto consented to a cease-and-desist order; the SEC imposed no civil penalty, citing its cooperation and remediation.

Founders and promoters. In April 2025 the SEC and federal prosecutors charged the founder of Nate, a shopping app, alleging he raised more than $40 million by telling investors the app completed purchases with AI and without human intervention, while contractors in a call center did the work and the automation rate was effectively zero. The charges are accusations, and he is presumed innocent unless proven guilty; the SEC’s civil case was stayed in June 2026. In September 2026 the SEC alleged that investors were promised returns from renting AI trading bots that did not exist and, in a companion case, from AI-generated trading signals issued through a fake platform.

Examinations. The SEC’s examination priorities for fiscal 2026, released in November 2025, say its staff “will review for accuracy registrant representations regarding their AI capabilities or AI”.

The 2024 cases that set the pattern. In March 2024 the SEC settled charges against two investment advisers, Delphia and Global Predictions, over false and misleading statements about their use of AI; without admitting or denying the findings, they agreed to civil penalties of $225,000 and $175,000. According to the SEC, Delphia agreed in 2021 to correct its statements, yet further misstatements continued through August 2023. In September 2024 Texas announced a settlement, filed in August, with Pieces Technologies, a health AI company, over the advertised accuracy of its clinical summaries; Pieces denied wrongdoing.

States. On October 14, 2026, several state attorneys general, including those of Nebraska and Arkansas, sued OpenAI alleging that it represented ChatGPT as safe while knowing of significant risks and limitations; these are allegations in pending cases.

Shareholders. Cornerstone Research counted 16 AI-related federal securities class actions in 2025 and 15 in the first half of 2026, against 7 in 2023. Its count includes companies that build AI infrastructure as well as those accused of misstating their AI, so it measures exposure, not misstatement alone.

What the orders say a claim needs

The orders are specific about what “substantiated” means.

• Evidence in hand each time the claim is made. The Workado order bars accuracy claims unless, each time, the company has competent and reliable evidence, based on standards generally accepted in the field, that the claim is true. The Air AI order requires written substantiation for earnings claims at the time they are made.
• Evidence a qualified person would accept. “Competent and reliable evidence” means tests, analyses or studies conducted and evaluated objectively by qualified people and generally accepted as yielding accurate results.
• The test records, kept. Workado must keep its testing protocols, the reasons for its metrics, its test data, the steps it took to keep training and test data apart, why the test data fit the real use, and its statistical analyses. The FTC said keeping that evidence guards against claims made without it, and that independent testing can be part of it.
• Metrics that say what they measure. Texas required Pieces to disclose what each accuracy metric means and how it is calculated, or to use an independent auditor.
• Someone responsible for the statement. In the Presto case, the SEC’s finding that no one was formally responsible for the accuracy of filings sat alongside the misstatements.

What would have prevented them

Are you at risk?

Ask these about every statement your organization makes about AI, inside or outside.

1. Could you show, today, the test behind each accuracy, speed or savings figure you publish about AI?
2. Was that test run on the kind of work your customers or investors will actually use it for?
3. Do you describe a supplier’s model or a human-assisted process as your own AI?
4. Does every figure come with what it measures and how it was calculated?
5. Is a named person responsible for the accuracy of AI statements in filings, pitches and marketing?
6. When the product changes, does someone check whether the claims about it still hold?
7. Would the claim survive a regulator asking for the evidence you had on the day you made it?

What to do next

1. Keep a register of AI claims: the wording, where it appears, who owns it and the evidence behind it.
2. Separate three kinds of claim. “We use AI” needs evidence that the system exists and is used as described; “AI improves the result” needs a measured comparison; “AI saves this much” needs a baseline, a period and a method.
3. Keep the test records the way the Workado order requires: protocol, data, metrics and why they fit the real use.
4. Re-check claims on a schedule. A claim that was true at launch can become untrue as the product changes, without anyone editing it.